def __init__(
self,
forward_func: Callable,
layer: Module,
device_ids: Optional[List[int]] = None,
) -> None:
r"""
Args:
forward_func (callable): The forward function of the model or any
modification of it
layer (torch.nn.Module): Layer for which attributions are computed.
Output size of attribute matches this layer's input or
output dimensions, depending on whether we attribute to
the inputs or outputs of the layer, corresponding to
the attribution of each neuron in the input or output
of this layer.
device_ids (list(int)): Device ID list, necessary only if forward_func
applies a DataParallel model. This allows reconstruction of
intermediate outputs from batched results across devices.
If forward_func is given as the DataParallel model itself,
then it is not necessary to provide this argument.
"""
LayerAttribution.__init__(self, forward_func, layer, device_ids=device_ids)
GradientAttribution.__init__(self, forward_func)
self.ig = IntegratedGradients(forward_func)
def __init__(self, model: Module, layer: ModuleOrModuleList) -> None:
"""
Args:
model (module): The forward function of the model or
any modification of it. Custom rules for a given layer need to
be defined as attribute
`module.rule` and need to be of type PropagationRule.
Model cannot contain any in-place nonlinear submodules;
these are not supported by the register_full_backward_hook
PyTorch API starting from PyTorch v1.9.
layer (torch.nn.Module or list(torch.nn.Module)): Layer or layers
for which attributions are computed.
The size and dimensionality of the attributions
corresponds to the size and dimensionality of the layer's
input or output depending on whether we attribute to the
inputs or outputs of the layer. If value is None, the
relevance for all layers is returned in attribution.
"""
LayerAttribution.__init__(self, model, layer)
LRP.__init__(self, model)
if hasattr(self.model, "device_ids"):
self.device_ids = cast(List[int], self.model.device_ids)
def __init__(
self,
forward_func: Callable,
layer: ModuleOrModuleList,
device_ids: Union[None, List[int]] = None,
) -> None:
r"""
Args:
forward_func (callable): The forward function of the model or any
modification of it
layer (torch.nn.Module or list(torch.nn.Module)): Layer or layers
for which attributions are computed.
Output size of attribute matches this layer's input or
output dimensions, depending on whether we attribute to
the inputs or outputs of the layer, corresponding to
attribution of each neuron in the input or output of
this layer. If multiple layers are provided, attributions
are returned as a list, each element corresponding to the
activations of the corresponding layer.
device_ids (list(int)): Device ID list, necessary only if forward_func
applies a DataParallel model. This allows reconstruction of
intermediate outputs from batched results across devices.
If forward_func is given as the DataParallel model itself,
then it is not necessary to provide this argument.
"""
LayerAttribution.__init__(self, forward_func, layer, device_ids)
def __init__(self, model: Module, layer: Module):
r"""
Args:
model (torch.nn.Module): The reference to PyTorch model instance.
layer (torch.nn.Module): Layer for which attributions are computed.
The size and dimensionality of the attributions
corresponds to the size and dimensionality of the layer's
input or output depending on whether we attribute to the
inputs or outputs of the layer.
"""
LayerAttribution.__init__(self, model, layer)
DeepLift.__init__(self, model)
self.model = model
def __init__(self, model: Module, layer: ModuleOrModuleList) -> None:
"""
Args:
model (module): The forward function of the model or
any modification of it. Custom rules for a given layer need to
be defined as attribute
`module.rule` and need to be of type PropagationRule.
layer (torch.nn.Module or list(torch.nn.Module)): Layer or layers
for which attributions are computed.
The size and dimensionality of the attributions
corresponds to the size and dimensionality of the layer's
input or output depending on whether we attribute to the
inputs or outputs of the layer. If value is None, the
relevance for all layers is returned in attribution.
"""
LayerAttribution.__init__(self, model, layer)
LRP.__init__(self, model)
def __init__(
self,
forward_func: Callable,
layer: ModuleOrModuleList,
device_ids: Union[None, List[int]] = None,
multiply_by_inputs: bool = True,
) -> None:
r"""
Args:
forward_func (callable): The forward function of the model or any
modification of it
layer (torch.nn.Module or list(torch.nn.Module)): Layer or layers
for which attributions are computed.
Output size of attribute matches this layer's input or
output dimensions, depending on whether we attribute to
the inputs or outputs of the layer, corresponding to
attribution of each neuron in the input or output of
this layer. If multiple layers are provided, attributions
are returned as a list, each element corresponding to the
attributions of the corresponding layer.
device_ids (list(int)): Device ID list, necessary only if forward_func
applies a DataParallel model. This allows reconstruction of
intermediate outputs from batched results across devices.
If forward_func is given as the DataParallel model itself,
then it is not necessary to provide this argument.
multiply_by_inputs (bool, optional): Indicates whether to factor
model inputs' multiplier in the final attribution scores.
In the literature this is also known as local vs global
attribution. If inputs' multiplier isn't factored in,
then this type of attribution method is also called local
attribution. If it is, then that type of attribution
method is called global.
More detailed can be found here:
https://arxiv.org/abs/1711.06104
In case of layer gradient x activation, if `multiply_by_inputs`
is set to True, final sensitivity scores are being multiplied by
layer activations for inputs.
"""
LayerAttribution.__init__(self, forward_func, layer, device_ids)
GradientAttribution.__init__(self, forward_func)
self._multiply_by_inputs = multiply_by_inputs
def __init__(
self,
model: Module,
layer: Module,
multiply_by_inputs: bool = True,
) -> None:
r"""
Args:
model (nn.Module): The reference to PyTorch model instance. Model cannot
contain any in-place nonlinear submodules; these are not
supported by the register_full_backward_hook PyTorch API
starting from PyTorch v1.9.
layer (torch.nn.Module): Layer for which attributions are computed.
The size and dimensionality of the attributions
corresponds to the size and dimensionality of the layer's
input or output depending on whether we attribute to the
inputs or outputs of the layer.
multiply_by_inputs (bool, optional): Indicates whether to factor
model inputs' multiplier in the final attribution scores.
In the literature this is also known as local vs global
attribution. If inputs' multiplier isn't factored in
then that type of attribution method is also called local
attribution. If it is, then that type of attribution
method is called global.
More detailed can be found here:
https://arxiv.org/abs/1711.06104
In case of Layer DeepLift, if `multiply_by_inputs`
is set to True, final sensitivity scores
are being multiplied by
layer activations for inputs - layer activations for baselines.
This flag applies only if `custom_attribution_func` is
set to None.
"""
LayerAttribution.__init__(self, model, layer)
DeepLift.__init__(self, model)
self.model = model
self._multiply_by_inputs = multiply_by_inputs
def __init__(
self,
forward_func: Callable,
layer: Module,
device_ids: Union[None, List[int]] = None,
) -> None:
r"""
Args:
forward_func (callable): The forward function of the model or any
modification of it
layer (torch.nn.Module): Layer for which attributions are computed.
Output size of attribute matches this layer's output
dimensions, except for dimension 2, which will be 1,
since GradCAM sums over channels.
device_ids (list(int)): Device ID list, necessary only if forward_func
applies a DataParallel model. This allows reconstruction of
intermediate outputs from batched results across devices.
If forward_func is given as the DataParallel model itself,
then it is not necessary to provide this argument.
"""
LayerAttribution.__init__(self, forward_func, layer, device_ids)
GradientAttribution.__init__(self, forward_func)
def __init__(
self,
forward_func: Callable,
layer: ModuleOrModuleList,
device_ids: Union[None, List[int]] = None,
multiply_by_inputs: bool = True,
) -> None:
r"""
Args:
forward_func (callable): The forward function of the model or any
modification of it
layer (ModuleOrModuleList):
Layer or list of layers for which attributions are computed.
For each layer the output size of the attribute matches
this layer's input or output dimensions, depending on
whether we attribute to the inputs or outputs of the
layer, corresponding to the attribution of each neuron
in the input or output of this layer.
Please note that layers to attribute on cannot be
dependent on each other. That is, a subset of layers in
`layer` cannot produce the inputs for another layer.
For example, if your model is of a simple linked-list
based graph structure (think nn.Sequence), e.g. x -> l1
-> l2 -> l3 -> output. If you pass in any one of those
layers, you cannot pass in another due to the
dependence, e.g. if you pass in l2 you cannot pass in
l1 or l3.
device_ids (list(int)): Device ID list, necessary only if forward_func
applies a DataParallel model. This allows reconstruction of
intermediate outputs from batched results across devices.
If forward_func is given as the DataParallel model itself,
then it is not necessary to provide this argument.
multiply_by_inputs (bool, optional): Indicates whether to factor
model inputs' multiplier in the final attribution scores.
In the literature this is also known as local vs global
attribution. If inputs' multiplier isn't factored in,
then this type of attribution method is also called local
attribution. If it is, then that type of attribution
method is called global.
More detailed can be found here:
https://arxiv.org/abs/1711.06104
In case of layer integrated gradients, if `multiply_by_inputs`
is set to True, final sensitivity scores are being multiplied by
layer activations for inputs - layer activations for baselines.
"""
LayerAttribution.__init__(self,
forward_func,
layer,
device_ids=device_ids)
GradientAttribution.__init__(self, forward_func)
self.ig = IntegratedGradients(forward_func, multiply_by_inputs)
if isinstance(layer, list) and len(layer) > 1:
warnings.warn(
"Multiple layers provided. Please ensure that each layer is"
"**not** solely solely dependent on the outputs of"
"another layer. Please refer to the documentation for more"
"detail.")
def attribute(
self,
inputs: TensorOrTupleOfTensorsGeneric,
target: TargetType = None,
additional_forward_args: Any = None,
interpolate_mode: str = "nearest",
attribute_to_layer_input: bool = False,
) -> TensorOrTupleOfTensorsGeneric:
r"""
Args:
inputs (tensor or tuple of tensors): Input for which attributions
are computed. If forward_func takes a single
tensor as input, a single input tensor should be provided.
If forward_func takes multiple tensors as input, a tuple
of the input tensors should be provided. It is assumed
that for all given input tensors, dimension 0 corresponds
to the number of examples, and if multiple input tensors
are provided, the examples must be aligned appropriately.
target (int, tuple, tensor or list, optional): Output indices for
which gradients are computed (for classification cases,
this is usually the target class).
If the network returns a scalar value per example,
no target index is necessary.
For general 2D outputs, targets can be either:
- a single integer or a tensor containing a single
integer, which is applied to all input examples
- a list of integers or a 1D tensor, with length matching
the number of examples in inputs (dim 0). Each integer
is applied as the target for the corresponding example.
For outputs with > 2 dimensions, targets can be either:
- A single tuple, which contains #output_dims - 1
elements. This target index is applied to all examples.
- A list of tuples with length equal to the number of
examples in inputs (dim 0), and each tuple containing
#output_dims - 1 elements. Each tuple is applied as the
target for the corresponding example.
Default: None
additional_forward_args (any, optional): If the forward function
requires additional arguments other than the inputs for
which attributions should not be computed, this argument
can be provided. It must be either a single additional
argument of a Tensor or arbitrary (non-tuple) type or a
tuple containing multiple additional arguments including
tensors or any arbitrary python types. These arguments
are provided to forward_func in order following the
arguments in inputs.
Note that attributions are not computed with respect
to these arguments.
Default: None
interpolate_mode (str, optional): Method for interpolation, which
must be a valid input interpolation mode for
torch.nn.functional. These methods are
"nearest", "area", "linear" (3D-only), "bilinear"
(4D-only), "bicubic" (4D-only), "trilinear" (5D-only)
based on the number of dimensions of the chosen layer
output (which must also match the number of
dimensions for the input tensor). Note that
the original GradCAM paper uses "bilinear"
interpolation, but we default to "nearest" for
applicability to any of 3D, 4D or 5D tensors.
Default: "nearest"
attribute_to_layer_input (bool, optional): Indicates whether to
compute the attribution with respect to the layer input
or output in `LayerGradCam`.
If `attribute_to_layer_input` is set to True
then the attributions will be computed with respect to
layer inputs, otherwise it will be computed with respect
to layer outputs.
Note that currently it is assumed that either the input
or the output of internal layer, depending on whether we
attribute to the input or output, is a single tensor.
Support for multiple tensors will be added later.
Default: False
Returns:
*tensor* of **attributions**:
- **attributions** (*tensor*):
Element-wise product of (upsampled) GradCAM
and Guided Backprop attributions.
If a single tensor is provided as inputs, a single tensor is
returned. If a tuple is provided for inputs, a tuple of
corresponding sized tensors is returned.
Attributions will be the same size as the provided inputs,
with each value providing the attribution of the
corresponding input index.
If the GradCAM attributions cannot be upsampled to the shape
of a given input tensor, None is returned in the corresponding
index position.
Examples::
>>> # ImageClassifier takes a single input tensor of images Nx3x32x32,
>>> # and returns an Nx10 tensor of class probabilities.
>>> # It contains an attribute conv4, which is an instance of nn.conv2d,
>>> # and the output of this layer has dimensions Nx50x8x8.
>>> # It is the last convolution layer, which is the recommended
>>> # use case for GuidedGradCAM.
>>> net = ImageClassifier()
>>> guided_gc = GuidedGradCam(net, net.conv4)
>>> input = torch.randn(2, 3, 32, 32, requires_grad=True)
>>> # Computes guided GradCAM attributions for class 3.
>>> # attribution size matches input size, Nx3x32x32
>>> attribution = guided_gc.attribute(input, 3)
"""
is_inputs_tuple = _is_tuple(inputs)
inputs = _format_input(inputs)
grad_cam_attr = self.grad_cam.attribute.__wrapped__(
self.grad_cam, # self
inputs=inputs,
target=target,
additional_forward_args=additional_forward_args,
attribute_to_layer_input=attribute_to_layer_input,
relu_attributions=True,
)
if isinstance(grad_cam_attr, tuple):
assert len(grad_cam_attr) == 1, (
"GuidedGradCAM attributions for layer with multiple inputs / "
"outputs is not supported.")
grad_cam_attr = grad_cam_attr[0]
guided_backprop_attr = self.guided_backprop.attribute.__wrapped__(
self.guided_backprop, # self
inputs=inputs,
target=target,
additional_forward_args=additional_forward_args,
)
output_attr: List[Tensor] = []
for i in range(len(inputs)):
try:
output_attr.append(guided_backprop_attr[i] *
LayerAttribution.interpolate(
grad_cam_attr,
inputs[i].shape[2:],
interpolate_mode=interpolate_mode,
))
except Exception:
warnings.warn(
"Couldn't appropriately interpolate GradCAM attributions for some "
"input tensors, returning empty tensor for corresponding "
"attributions.")
output_attr.append(torch.empty(0))
return _format_output(is_inputs_tuple, tuple(output_attr))
